Agentic Commerce: What Changes When the Buyer Is an LLM

A retail website is a UI optimised for a human in a hurry: hero banner, search box, faceted filters, cart drawer, three-step checkout. Every pixel assumes a pair of eyes. The new buyer doesn't have eyes. It has a prompt and a tool registry.

What "agentic commerce" actually means

When a customer asks ChatGPT for "a black wool coat under €300, ships to Germany," and the LLM calls a tool that lists matching SKUs from your catalog, you're already in agentic commerce. The website never opens. The customer never sees your stock photography. They see one paragraph the model wrote and a buy button.

The merchant-side shift

Three things change for the merchant:

1. Catalog goes from HTML to structured data. The model doesn't read your product description copy; it reads the schema. Title, brand, attributes, stock, price, currency, return policy — every field that used to be optional is now load-bearing. Missing availability and your SKU is invisible.
2. Storefront-as-Tool. Your store ships an MCP server (or ACP endpoint) exposing search_products, get_product_details, add_to_cart, create_checkout_session. The customer's AI calls those tools. You don't run the conversation; you run the inventory of the conversation.
3. Brand happens earlier. The customer doesn't browse your storefront, so you can't recover with great photography or microcopy. Whatever the model says about your brand at the moment of comparison is what your brand is. That signal lives in the structured data + the model's training, not your website.

What still works

• Real-world fulfilment. AI picks the SKU; humans still pack the box. Anything tied to physical reality (warehouses, shipping speed, return windows) is unchanged.
• The merchant of record. Payment processing, tax calculation, refund policy — these stay where they were. The model is a discovery and decision layer; settlement is identical.

What stops working

• Funnels. "Visitor → cart abandoner → email retarget" assumes a website session. AI sessions don't return; the customer asked about coats once, the model answered once. You either won at the moment of comparison or you didn't.
• Discount codes as marketing. A code makes sense when a human can type it. An LLM that sees BLACKFRIDAY30 in your sitemap will use it for everyone, every time. Discount logic moves into rules the model evaluates: "for repeat buyers, X; otherwise Y."

The two-year forecast

A chatbot bolted onto your existing site doesn't win agentic commerce. The catalog wins it — when you treat it as a programmable surface (versioned, schema-typed, idempotent on writes) and demote the storefront UI to a separate read view.